[eshulman2]

Summary

Add safe user-defined workflow nodes as inline templates or active project-published Node definitions. New runs resolve project references to exact revisions and pin the resolved template in the workflow artifact.

• decision-v1 evaluates ordered typed predicates and emits a declared outcome.
• agent-assessment-v1 sends only selected workflow inputs to one structured, tool-free model call and records a bounded summary.
• Conditional edges can read node.<step>.outcome, including any/all branch scopes at join: all.
• Add project-scoped node publication, activation, rollback, catalog, and CLI inspection, plus an artifact-review example.
• Include template content in workflow manifest comparison and persist branch-keyed outcome records in checkpoints.

Verification

• pytest -q tests/unit/workflow tests/integration/workflow/test_definition_publication.py — 1093 passed, 1 skipped
• ruff check — passed
• forge node validate docs/examples/workflow-nodes/artifact-review-node.yaml — passed

Stacked on #347 (feature/conditional-workflow-edges).

[eshulman2]

Demo path:

1. Inspect and validate the example: forge node catalog and forge node validate docs/examples/workflow-nodes/artifact-review-node.yaml.
2. Publish it for a test project: forge node publish DEMO docs/examples/workflow-nodes/artifact-review-node.yaml, then forge node activate DEMO artifact-review 1.
3. Add a review_artifact step after specification generation in a project workflow, using node: {source: project, name: artifact-review} and a case for {fact: node.review_artifact.outcome, op: equals, value: ready}. Route ready to task generation and otherwise back to specification revision. The exact YAML snippet is in docs/examples/workflow-nodes/README.md.
4. Run a DEMO ticket through the workflow with a specification. Inspect its checkpoint: node_results.review_artifact:main contains the declared outcome, short summary, and attempt. Publish and activate a second node revision, then verify the existing run keeps its pinned template while a new run uses revision 2.

The assessment node requires a configured model and a nonempty specification; missing input or malformed model output blocks the run.